Last month, Anthropic quietly released a model called Claude Mythos Preview. The thing was terrifyingly capable.
The UK AI Safety Institute put it through its paces. They found it could autonomously execute a full corporate cyberattack chain—32 steps that human experts estimated would take 20 hours. It completed the whole thing on its own. On expert-level CTF challenges, it hit a 73% success rate. That’s the kind of number that makes security researchers lose sleep.
Anthropic’s response? Lock it down. Mythos Preview was only made available to a consortium of about 40 tech companies under Project Glasswing, specifically for critical infrastructure security assessments. The company stated clearly they had no plans for public release. Too capable, too risky—keep it behind closed doors.
But here’s the catch: Anthropic never published any technical details about Mythos. No paper, no architecture diagram, no blog post explaining how it works. The entire thing was a black box.
Then a 22-year-old open-source developer named Kye Gomez decided he wasn’t going to wait for Anthropic to share.
He spent weeks digging through publicly available academic papers—not from Anthropic, but from other labs working on related ideas. He pieced together clues, made assumptions, and built what he believes Mythos might look like under the hood. The result is a project called OpenMythos, a pure PyTorch implementation of a hypothetical Mythos architecture.
The project hit 7,000 stars on GitHub in four days. It’s now sitting at over 11,000. The Issues section? A battlefield. Some call it genius. One open issue is literally titled “This is still dumb.”
So what exactly did OpenMythos accomplish?
Traditional large language models—GPT, LLaMA, the whole crowd—scale by stacking layers. 100 layers? Not enough. 200? Still not enough. 400? Getting closer. Each layer has its own independent set of parameters, and the parameter count balloons accordingly.
OpenMythos takes a completely different approach: instead of stacking more layers, it reuses the same set of weights multiple times.
Think of it this way. A traditional model reads a book page by page: you turn the page and move on. OpenMythos is like having someone read the same paragraph over and over, each time understanding it a little deeper. The depth of reasoning isn’t determined by how many parameters you have, but by how many times you let the model “think.”
The numbers are compelling. A 770M parameter recurrent model can match a 1.3B traditional Transformer. Same performance, nearly half the parameters. That’s not just efficiency—it’s a fundamental shift in how we think about scaling.
The architecture itself breaks down into three segments: Prelude, Recurrent Block, and Coda.
The Prelude and Coda are standard Transformer layers that process once. The Recurrent Block is where the magic happens—the same set of weights runs T times.
Three features stand out.
First, iterative deepening. Each pass through the Recurrent Block mixes the current state with the original input. It’s not just repeating the same calculation; each round builds on the previous one for deeper reasoning. The most impressive part is the depth extrapolation capability: train on 16 rounds, then at inference time you can run 24, even 32 rounds, and the model generalizes to depths it’s never seen before. Simple questions? Fewer rounds. Hard questions? More rounds. You don’t change the model, just a parameter.
Second, MoE + MLA for efficiency. The Recurrent Block uses a Mixture of Experts with DeepSeekMoE’s fine-grained routing mechanism. Different depths can activate different expert subsets, meaning the same weights perform different functions at different steps. For attention, you have two options: Multi-Latent Attention from DeepSeek-V2 (which shrinks KV cache by 10–20x) or Grouped Query Attention with Flash Attention 2 support. There’s also an Adaptive Computation Time mechanism that lets simple tokens exit early while hard ones keep computing—all in the same batch.
Third, mathematically guaranteed training stability. Recurrent Transformers have a notorious problem: training instability. Run the same weights repeatedly and gradients explode or vanish. Many attempts have died on this hill. OpenMythos injects LTI—Linear Time-Invariant system discretization. The matrix A is constructed so its spectral radius is strictly less than 1. No matter your learning rate, training is mathematically stable. All calculations happen in log-space with clamping to prevent float32 overflow. This isn’t tuning—it’s proof.
Now for the honest part.
OpenMythos currently has no trained weights, no published benchmarks, no actual inference outputs. It compiles. The architecture is correct. But it hasn’t produced any results yet.
The GitHub Issues reflect the split. Some users are running third-party benchmarks. Others are suggesting Flash Attention and mixed-precision optimizations. And some are just yelling.
One user requested official benchmarks. No response yet.
So OpenMythos’s position is clear: it’s a code implementation of a hypothetical architecture, not a ready-to-use product. It’s saying: if Claude Mythos really is a recurrent depth Transformer, here’s what that might look like.
That in itself is remarkable. In open-source AI, most projects replicate existing papers. OpenMythos replicates an architecture that was never confirmed to exist.
To try it yourself:
pip install open-mythos The project comes with seven preset configurations from 1B to 1T parameters. You can switch between MoE and standard attention backends. Training scripts are in the training/ directory, using PyTorch FSDP for distributed training, default dataset FineWeb-Edu.
Here’s the bigger picture. OpenMythos pushes the AI scaling debate from “how many parameters do we stack” to “how many rounds of computation do we allow at inference time.” Whether or not Claude Mythos actually uses this architecture, the direction itself is worth paying attention to.
770M matching 1.3B saves more than parameters—it lowers training costs and deployment barriers. And right now, we have a specific, testable hypothesis. That’s something you can work with.